The Kainoki Treebank front page

Table of contents

• Chapter 1 Annotation overview
• Chapter 2 Particles/Postpositions (P)
• Chapter 3 Phrase layers
• Chapter 4 Predicates
• Chapter 5 Clause layers
• Chapter 6 Constructions
• Chapter 7 Ambiguous forms

Chapter 1
Annotation overview

1.1   Introduction

This manual details an annotation scheme for parsing contemporary Japanese. Syntactic structure is represented with labelled parentheses adapting the format of the Penn Treebank (Bies et al. 1995). More particularly, the Penn Historical Corpora scheme (Santorini 2010) has informed the ‘look’ of the annotation. This includes:

• adoption of the CorpusSearch format (described in section 1.3.1 below) as the underlying encoding,
• not having any explicit verb phrase structure (although verb phrase structure is implicitly present when there are interpretive consequences),
• the use of IP, ADVP, NP, and PP tag labels,
• the presentation of phrase conjunction structure with CONJP layers, and
• the marking of function for all clausal nodes and all clause level constituents.

Annotation practice is also heavily inspired by the SUSANNE scheme (Sampson 1995) regarding adoption of construction analysis, and most in particular for the methods of representing complex expressions with subordination and/or coordination. The annotation also contains plenty that is innovative, including information to resolve both inter and intra sentential anaphoric dependencies from a discourse perspective.

    Annotation is carried out with a general text editor on text files.

    Annotation practice strives first for observational adequacy. The aim is to present a consistent linguistic analysis for each attestation of an identifiable linguistic relation or process. Relations and processes are treated uniformly as much as possible, and their treatments are detailed in this documentation. Setting aside the issue of whether the system of description is theoretically correct, the practice is to render lexical and functional items, parts of speech, constituents of various categories and functions, and constructions defined by combinations of properties, in such ways as to be unambiguously identifiable. This documentation sets out basic principles both for the annotator (assigning segmentations, tags, and structural positions) and for the user (searching for classes of items, categories, and relationships between these). Examples of suitable tools for searching the parsed data without requiring any modification to the data include: CorpusSearch (Randall 2009) and Tregex (Levy, and Andrew 2006).

    The current annotation also aims to offer syntactic analysis that can serve as a base for the subsequent generation of meaning representations using the methods of Treebank Semantics (Butler 2015). To this end, extra disambiguation information is added to feed the calculation of semantic analyses from the syntactic annotation. One example of this is seen in the different specifications of clause linkage (i.e., different types of non-final clauses). The annotation has the -SCON tag extension to identify clauses integrated with subordinate conjunction. Subordinate clause status influences the distribution of empty subject positions within such clauses and the antecedence relationships these positions have with respect to upstairs arguments (according to an antecedent calculation called ‘control’). These cases are contrasted with coordinate clause linkage, also identified with a tag extension: -CONJ (coordinating conjunction). Status as a coordinate clause influences the distribution of arguments shared between that clause and one or more other clauses (according to an antecedent calculation called ‘Across the Board extraction’ (ATB)).

    With these calculations in place, most antecedent relations in Japanese can be accurately determined without resorting to annotation with overt indexing, provided that the distinction between subordination and coordination is properly annotated. The practice provides a robust basis for calculating semantic dependencies, a simplified annotation scheme that is descriptively adequate, and a set of constraints on the distribution of null positions which have interesting consequences with respect to, for example, the placement of null elements.

    The purpose of the rest of this chapter is to introduce broad aspects of the annotation scheme, while subsequent chapters will focus on particular topics.

1.2   Tags

Tag labels are either word class tags (e.g., N=noun, P=particle, ADV=adverb), phrase level categories with minimally a basic label to indicate the form of the constituent (e.g., NP=noun phrase, PP=particle phrase, ADVP=adverb phrase), or clause level categories (e.g., IP-MAT=matrix clause, CP-FINAL=matrix clause with sentence final particle(s)). Frequently, label extensions (separated by a hyphen) are added to labels to indicate function (e.g., P-ROLE=role particle, NP-SBJ=subject noun phrase). In most cases there is one label extension, but more is possible (e.g., CP-QUE-OB1= question used as object, IP-ADV-SCON-CND=conditional clause).

    All tag labels used in this schema are listed below in Tables 1.1–1.5.

Table 1.1: Word class tags

Table 1.2: Phrase layer tags

Table 1.3: Clause layer tags

Table 1.4: Function tags

Table 1.5: Other tags

1.3   General parsing principles

The annotation scheme represents syntactic structure with labelled parentheses. All open parentheses have an associated label, representing non-terminal nodes in a tree.

1.3.1   Format of the parsed data

All corpus files are formatted to be compatible with the CorpusSearch program (Randall 2009; http://corpussearch.sourceforge.net/CS-manual/YourCorpus.html). Under CorpusSearch format, every tree has a ‘wrapper’. A wrapper is a pair of unlabelled parentheses surrounding the tree content together with an ID node. Furthermore, to support the on-line corpus interface, the ID node must contain a string that begins with a number corresponding to the number of the tree in the file which is followed by an underscore (‘_’) and in turn followed by the name of the file (minus the ‘.psd’ extension). Any other string material may then follow the filename, provided there is an intervening semicolon (‘;’).

1.3.2   Terminal nodes

Words are terminal nodes that project a node with a word class label. Null elements (zero pronouns, traces, etc.) that aren't in the text itself but are introduced to describe structure are placed directly under phrase-level nodes, and so have no word class node projection. Null elements have terminal nodes that begin and end with ‘*’. See sections 1.5 and 1.4 for details about null elements. Terminal elements that are words and so under word class nodes are surrounded by ‘*’ when they are insertions made to the data from the stages of annotation, as in the recovery of content for instances of particle omission (see section 2.14).

1.3.3   Segmentation and part-of-speech annotation

The policy for segmentation and part-of-speech labelling follows the principle of using terminal nodes that are as large as possible, but not so large as to incorporate into purely lexical elements other elements with functional roles. Such a policy corresponds closely with the LUW (Long Unit Word) standard of the Corpus of Spontaneous Japanese (CSJ; Maekawa 2003) and the Balanced Corpus of Contemporary Written Japanese (BCCWJ; Maekawa et al. 2014). A LUW (Long Unit Word) is composed of at least one SUW (single-morpheme Short Unit Word), which corresponds to an entry in the UniDic dictionary (Den et al. 2008), but complex LUWs containing more than one SUW are common.

    The chunking obtained with LUW analysis is not limited to complex nominal expressions or complex predicates: Heterogeneous strings that appear to have undergone grammaticalisation (e.g., some formal noun/particle pairs, some complex modal expressions, etc.) are chunked as well. Complex LUWs are usually incorporated just as single segments. For example, numerals are analysed digit-by-digit into component SUWs, but the parsed annotation chunks these into a single segment according to the LUW containing them.

    While the LUW chunking of the BCCWJ and CSJ is intended to identify units with significance in syntax, the information is not always rich enough to generate immediate constituency trees approaching descriptive adequacy for syntax. Depending on the circumstances, SUW analysis may not be sufficiently fine grained, making it necessary to split the SUW into more than one segment. For example, the volitional form of a verb may be analysed into a combining stem and a volitional morpheme: (VB 結ぼ) (AX う). Conversely, a series of LUWs may need to be concatenated under one terminal node label (an instance of further chunking). For example, a series of LUWs that together form a complex proper noun may be chunked into one segment. Furthermore, some finer distinctions in morphological analysis that have no consequence for syntax (e.g., the distinction between personal names and place names) are sometimes ignored, while other distinctions deemed important (e.g., sorting instances of items that share the same phonological form into more than one part of speech according to their grammatical function) are introduced. This is a consequence of aiming to expose the basic functional structure of the language, while keeping the structure fairly flat and easily searchable.

    The policy is to chunk as large as possible in the automatic parse, and this is the form of segments that annotators initially see. But when there is clearly some constituency in a string that must be expressed by structure, or when there is a need to indicate the semantic effects of structure, chunking may have to be undone. To exemplify the former situation, consider the morpheme 中 (chuu). Following a ‘verbal noun’ like 旅行, 中 is analysed as a nominalising suffix by UniDic and grouped together with a preceding string to form a LUW. This analysis is appropriate for a situation in which 旅行 takes a noun modifier, as in (1.1).

(1.1)

彼 は 旅行中 にして 、 病 に 倒れ た 。

But in a different context 旅行 may appear with arguments or adverbial elements, and 中 is better analysed as a noun, as in (1.2).

(1.2)

佐藤さん は 海外 を 旅行 中 だ 。

    Another scenario where chunking may have to be undone involves complex particles. UniDic chunks many verb-particle combinations into complex particles. Note how in the string にしたがって in (1.3) the part corresponding to the verb 従う has been bleached of its semantics.

(1.3)

「 地 は 生き物 を 種類 にしたがって いだせ 。

While the UniDic analysis is frequently correct, there still appear instances where the part corresponding to the verb 従う should be treated as a full-fledged verb, as in (1.4).

(1.4)

モーセ は 主 の 命 に したがっ て 、 パラン の 荒野 から 彼ら を つかわし た 。

In a scenario such as (1.4), the annotator must split the segment and relabel its parts as necessary.

1.3.4   The schema of phrase constituents

Phrase structure is roughly describable as in (1.5):

(1.5)

    Words that project phrases are phrase heads (e.g., N, P, ADV, etc.). Following (1.5), phrase heads are immediately dominated by phrase nodes (e.g., NP, PP, ADVP, etc.), so modifiers and complements of the phrase head generally appear as phrases or clauses sister to the head. Intermediate levels of structure in the sense of X' theory (N', ADV', etc.) are not represented explicitly in general.

1.3.5   Endocentric structure and exceptions

In general the head (N, P, ADV, etc.) is overt and matches the category of the immediately dominating phrase level (NP, PP, ADVP, etc.).

(1.6)

(1.7)

    There are cases in which the head is not overt or does not match the category of the phrase level. There are various scenarios in which the category of the word level constituent does not match the category of the phrase that it projects.

• The head has a more specific label than its general category label (as can be seen in the way that the pronoun (PRO), as a subclass of nouns (N), heads an NP in (1.8)).

(1.8)

• More specific heads of NP include proper noun (NPR), quantifier (Q), pronoun (PRO), indeterminate pronoun (WPRO), foreign word (FW), mentioned expression (N-MENTION), and intermediate nominal (NML). As an exceptional case, NUMCLP (itself a phrasal category) always projects an NP.
• As inflecting elements, predicates regularly head IP, but predicates can have core elements of various types: VB; ADJI; ADJN+AX; NP-PRD+AX; all project a clause (IP).
• In numeral-classifier phrases (NUMCLP) the numeral (NUM) and the classifier (CL) are mutually dependent, and while the rightmost element (normally a CL) is the head, both words are directly dominated by a syncretic category NUMCLP.
• ADVP is frequently headed by ADJN+AX in the infinitive inflection, and sometimes by ADJI in the infinitive inflection.
• A prenominal phrase (PNLP) is allowed to dominate any variety of category, in any arity.
• The category CONJP can be headed by either a P (typically P-CONN), a CONJ, or a ‘bare’ phrasal category.

    There are cases where the head of the phrase has been elided. This can be seen in the lack of a predicate in the IP-ADV-CONJ of (1.9) — a right-node raising construction.

(1.9)

仙台フィルハーモニー管弦楽団 の 第２７９ 回 定期演奏会 は 、 指揮 に ヘンリク・シェーファー 、 ピアノ に 萩原麻未 を 迎え ます 。

1.3.6   Basic clause structure

A matrix clause is labelled IP-MAT. The pairing of a subject (either overtly expressed in the annotation or not) and a predicate together projects a clause (IP, or ‘inflectional phrase’). Typically, a clause has one main predicate such as a verb (VB), イ-adjective (ADJI), ナ-adjective (ADJN) with copula (AX), or nominal predicate (NP-PRD) with copula (AX). The annotation does not invoke a verb phrase level, and so IP structure is typically flat, with clause level constituents represented at the same level as main predicate material.

• Verb as main predicate:

(1.10)

花子 が 泣い た 。

• イ-adjective as main predicate:

(1.11)

この 料理 は おいしい 。

• ナ-adjective together with copula as main predicate:

(1.12)

キム は 有能 だ 。

• NP-PRD together with copula as main predicate:

(1.13)

メアリー は 女性 で ある 。

    In principle, phrases (rather than words) combine with predicates under IP to compose clauses with complex meanings. In addition to core predicative material (verb, イ-adjective, ナ-adjective + copula, nominal predicate + copula), a small number of other word level constituents can be immediately dominated by IP. These are: coordinating conjunctions (CONJ), single-word interjections (INTJ), auxiliary verbal morphology (VB0, VB2, PASS, AX, etc.), sentential negation (NEG), modals (MD), some formal nouns (FN), and particles (P).

(1.14)

電話番号 を 聞き 違え た か 、 または もう この 電話 は 使わ れ て い ない の だろう 。

1.3.7   Fragments (FRAG)

When there is a series of one or more phrases which does not form an IP, the construction is a fragment and the member phrase(s) can be placed under FRAG.

(1.15)

まったく 。

(1.16)

＝ ２０１１ 年 ４ 月 、 遠野市

(1.17)

おそらく 、 ある 観念 も 。

Single NP or PP elements used as exclamative utterances fall into this scenario.

(1.18)

泥棒 ！

This policy applies to imperative utterances formed on こと as well:

(1.19)

この 第6 作 について は 『 仮面ライダー （ スカイライダー ） 』 を 参照 の こと 。

Note that FRAG can appear under CP-THT and CP-QUE:

(1.20)

糞坊主め と はがみ を し た 。

(1.21)

何 の ？

1.3.8   Parenthetical layers (PRN)

The node label PRN is used for elements bounded by brackets or for an element in an appositive relationship with another element. Dashes and parentheses (PUL and PUR) that set off a PRN will be contained within the PRN node.

    Instances of PRN contain a variety of different projections and cover a wide range of various functions some of which are, contrary to their parenthetical nature, semantically significant. In particular, they can be mere side remarks to, functional equivalents of, or even quantifiers of, those which they modify.

(1.22)

○ 国務大臣 （ 三塚博君 ）

(1.23)

津波 を 防御 し た 仙台東部道路 の 内陸側 に 立地 する ＝ 仙台市 若林区 荒井東

    Among the instances of appositive PRN, the following sorts can be seen:

• Frequently a PRN will be next to either a common noun (N), a proper noun (NPR) or a pronoun (PRO), adding supplementary information to the head of the NP dominating them both.

(1.24)

東北 の 空 の 玄関口 ・ 仙台空港 は 、 １６ 年 ３ 月 の 民営化 方針 が 正式決定 し た 。

(1.25)

我々 が ―― 我々 日本帝国人民 が 偉い か 、

• A PRN can also break into speech acts, containing an insertive sentence (an IP-MAT or a CP) and then being an element of another clause.

(1.26)

答える べき 質問 は -- もう 時間 です ね -- 毎日 の 生活 の 中 で どう すれ ば より 多く の 時間 を フロー の 状態 に できる か

• A PRN immediately following common noun (N), a proper noun (NPR), or a pronoun (PRO) may contain an NP headed by a reflexive pronoun (for details about reflexive pronouns, see section 3.2.2 ).

(1.27)

―― それ は 支配人 自身 だっ た 。

(1.28)

そして 私 は その 中 に 現実 の 私 自身 を 見失う の を 楽しん だ 。

• A PRN immediately following a common noun (N), a proper noun (NPR), or a pronoun (PRO) may contain a quantificational NP headed by a numeral classifier phrase (NUMCLP), a quantifier (Q), or a noun/quantifier word (QN), which quantifies those preceding nominal head. In this case, the quantificational NP inside the PRN must be marked by ;*, as in (1.29) below. (see section 3.3.5.2 for details).

(1.29)

社員 全員 が 昇進 を 目指し て いる 。

(1.30)

うち 多賀城市 は 対象 の 農地 ９７ ヘクタール の 工事 が 全て 完了 し た 。

• A PRN can embrace a CP-THT or a CP-QUE which complements some postceding NP (e.g., headed by 発言, 質問, 答え, 動機, 傾向, etc.) which semantically requires a statement or a question and is allgedly drifted apart from that NP. In this case, the PRN and its trace will be *ICH*-marked (section 1.4). For details see section 6.18.3 and 6.17.2.

(6.58)

芸術界 で 今 何 が 起こっ て いる か 動向 が 分かる と さ れる もの でし た

(1.32)

真宗 の 寺院建築 に は 他 に も 内陣 に比べて 外陣 が 広い など 、 他宗 に 見 られ ない 特徴 が ある 。

1.3.9   Punctuation

As a general rule, punctuation is attached at the highest possible level, with the exception that punctuation cannot occur under multi-sentence.

(1.33)

希望者 は 、 住所 、 氏名 、 および 、 電話番号 を 記入 し て 下さい 。

Note especially that when an overt coordinating conjunction follows an IP-ADV, punctuation intervening between the IP-ADV and the following coordinating conjunction is allowed.

(1.34)

彼女 は 一流 の ピアニスト で 、 かつ 優れ た 随筆家 だ 。

    As is already mentioned in section 1.3.8, another exception is dashes and parentheses (PUL and PUR) setting off a parenthetical element (PRN). They are contained within the PRN node.

1.3.10   Metadata (META)

Occasionally texts include metadata such as stage directions or identification of the person speaking. This kind of marginalium is placed under the tag META.

(1.35)

週末 は 人並み が 絶え ない ＝ 昨年 １１ 月

1.4   A null element that must be indexed

There is one null element that must be indexed:

Table 1.6: Null element with indexing

*ICH* (abbreviating ‘Interpret Constituent Here’) is used as a trace marking discontinuous structures, e.g., rightmost ‘afterthought’ (see section 6.21.4), long-distance scrambling, or other remote relationships that cross phrase levels but are neither calculated (by control, ATB, relative clause extraction), nor specified by sort information (see section 1.5.6 for discussion). Indexing works by adding an index number to the label of the remote (overt) constituent, creating a dummy constituent of the same category in the place where the remote constituent should be interpreted, and adding as its child an *ICH* which shares the same index number. For example, in the following example, a post-posed particle phrase corresponds to an indexed *ICH*. Note that extensions indicating grammatical function (e.g., -SBJ) are marked on the node label dominating *ICH* and not on the node label of the remote constituent.

(6.106)

美しい 街 です よ 、 神戸 は 。

    It is possible for a single element to be coindexed with multiple instances of *ICH*, as in examples (1.37) and (1.38):

(1.37)

2012 年 末 の 男性平均寿命 80.18 歳 、 女性平均寿命 84.67 歳 、 戸籍人口平均寿命 82.41 歳 。

(1.38)

訪問診療 の 患者 は 月 ２０ 人 程度 おり 、 定期的 な 巡回診療 は 町外 の 仮設住宅 で 月 ３ 回 __、 離島 ___の 出島 で 月 ２ 回 、 江島 で 月 １ 回 を それぞれ こなす 。

    Note that cases in which constituents may have been reordered leftward but remain at the same clause level (e.g., topicalisation, short-distance scrambling, etc.) are not indicated.

(1.39)

この 服 は 太っ た 人 でも 着 られ ます 。

    By contrast a rightmost ‘afterthought’ (see section 6.21.4) is placed at a different level from the clause which should contain a coindexed *ICH*, thereby ensuring the main predicate has rightmost placement:

(1.40)

「 見よ 、 神 の 小羊 」 。

    Floating noun-complement clauses adjoining to upper ones also falls under the *ICH*-treatment. The caveat is that these floating clauses are covered by a PRN projection, which it is that undergoes indexicalization. (See also sections 1.3.8, 6.18.3 and 6.17.2)

(1.41)

先生 は また 玄関 の 前 に 戻っ て 、 「 前 へ なら え 。 」 と 号令 を かけ まし た 。

1.5   Null elements without indexing

While the single null element of the previous section requires indexing, the null elements of this section are assigned nodes in the annotation scheme without indexing. These unindexed null elements include null expletives, zero pronouns of various kinds, and trace markers of relative clauses. The full list of unindexed null elements used is shown in Table 1.7 along with specifications for their use.

Table 1.7: List of null elements without indexing

    Note that gaps for controlled subject positions (controlees — see section 5.3), and gaps formed by Across the Board extraction (ATB — see section 5.4) are not assigned nodes in the annotation.

1.5.1   Zero element

A zero element is used when no claims are wished to be made about content but still material is required to support the projection of a constituent. The notable case where this may be required is with the logical subject of a passive clause with (NP-LGS *). This makes no claim about the content of the logical subject besides its need to be present to satisfy the requirements of the passive construction.

(1.42)

財政 は 借金まみれ で 不利益 の 分配 を 迫ら れる 。

1.5.2   Null expletive

Sentences that appear not to have subjects are supplied with a null element subject (NP-SBJ *exp*). This includes:

• weather predicates:

(1.43)

明日 は 寒く ない だろう 、

• and special types of clauses formed entirely on nominal predicates.

(1.44)

非常 に 落胆 し た 石森 だっ た が 、 50 枚 以上 の デザイン画 を 描い た 。

• If an utterance can be construed as a complete assertion about a state of affairs in the world, a clausal analysis is favored and an expletive subject is supplied where an interpretable subject is not retrievable from context:

(1.45)

「 火事 だ ！ 」

1.5.3   Zero pronoun with generic impersonal reference

As (1.46) demonstrates, *arb* is used for generic impersonal reference.

(1.46)

論文 を 書く の は とても たいへん です 。

1.5.4   Other zero pronouns

Like an overt pronoun, *pro* may refer to an entity in the discourse context, or may take as an antecedent an element in the same sentence. *hearer*, *hearer+pro*, *speaker*, *speaker+hearer* and *speaker+pro* are more specified versions of *pro*, and are preferred over *pro* when appropriate.

(1.47)

何 を 買っ て あげ よう か 。

(3.2)

___「 先生 、 もう お忘れ です か ？ 」

    In examples (1.47) and (3.2), zero pronouns mark non-controlled covertly expressed core grammatical roles (NP-SBJ, NP-OB1, NP-OB2), i.e., arguments that are obligatory for interpreting the clausal predicate, but which have null expression and which cannot be reconstructed from either control relations (see section 5.3) or antecedent relations arising from coordination structures, specifically ATB extraction (see section 5.4).

1.5.5   Traces in relative clauses

A trace *T* indicates the existence of an empty position in a relative clause modifying the head of a noun phrase that corresponds to the empty position. For more details, see section 3.2.6.1. Example (1.49) includes a trace for a subject (NP-SBJ *T*) to link the head (N 人) with the subject grammatical role inside the relative clause.

(1.49)

わかっ た 人 は 手 を あげ て ごらん なさい 。 」

    Note that traces are not limited to core grammatical roles: Traces can indicate the presence of ‘gaps’ in adjunct positions as well as in argument position. In this sense traces contrast with zero pronouns like (NP-SBJ *pro*), which correspond only to arguments with core grammatical roles. For details, about adjunct traces, see section 1.7.3.

    An IP-REL that doesn't contain a coordinated clause can have only one local trace. For an IP-REL that contains a coordinated clause, there can be at most one trace per conjunct, although there is no constraint on the grammatical functions of these traces, and no requirement that all the conjuncts contain a trace. For example, in (1.50) the first conjunct in the coordinated IP-REL has a trace in object position, while the second conjunct has a trace in subject position.

(1.50)

彼女 は 皆 が 愛し そして 皆 を 愛する 人 です 。

The second trace in (1.50) is placed to the right of the first conjunct. Otherwise, Across the Board extraction would require co-referential positions with identical grammatical roles in both conjuncts.

    One more thing to note about traces: If an overt pronoun or *pro* in a relative clause is co-referent with the trace *T*, this relationship must be indicated by adding sort information (see section 1.8 for details) not on the trace, but on the NP that the modified noun phrase head projects, as seen with example (3.33).

(3.33)

有間皇子 が どう し て も そこ から 逃れる こと の でき なかっ た 悲運

1.5.6   The position of unindexed null elements

With null elements not being visible, string placement for a null element is often not obvious. Position is therefore determined as follows:

• Almost all relative clause traces (*T*) appear in constituent-initial position, but observe remarks on trace placement given earlier in section 1.5.5.
• An expletive (*exp*) is IP-initial, with the exception that a trace with no other constraints on its positioning is placed before an expletive.
• In general, zero pronouns (*arb*, *exp*, *pro*, *speaker*, *hearer*, *speaker+hearer*, etc.) are placed IP-initial, so before conjunctions, and left-dislocated elements, but after traces (*T*) with no other constraints on their positioning, and expletives (*exp*).

Note that antecedent calculation has an influence on the placement of zero pronouns, as is discussed in the remainder of this section.

(1.52)

よそ から 、 もらっ た お酒 が 二 升 あっ た 。

(1.53)

この ビル は 、 現在 の 耐震基準 に 照らし合わせる と 、 強度 が 不足 し て いる ．

    The calculation of antecedents for empty positions can affect the placement of zero pronouns. For example, the calculation observes a rule for co-reference into coordinate clauses called ‘Across the Board’ extraction (ATB). Specifically, as Figure 1.1 illustrates, any overt category x in IP_n must be an antecedent to one empty position (preserving its grammatical role) in each of coordinate clauses IP_{1, 2...n-1} disambiguated with CONJ provided that x is sister to and precedes IP_n-1, and provided that IP_n-2 is the first element in IP_n-1, and provided that IP_n-3 is the first element in IP_n-2, and so on. This requires that each of IP_{1, 2...n-1} ‘inherit’ the antecedent relation from the category in IP_n. Placement of a zero pronoun in IP_n may be decided by whether that zero pronoun is an antecedent for all non-final conjuncts (in which case the zero pronoun is leftmost).

Figure 1.1: ‘Across the Board’ extraction (ATB).

(1.54)

「 和国 の 教主 」 として 尊敬 し 、 観音菩薩 の 化身 として 崇拝 し た 。

    In (1.55), the zero pronoun subject of the final clause (the matrix clause) is the antecedent of an empty position in the non-final clause (the precedence of the coordinated clauses). In the non-final clause a null pronoun for a primary object is placed in initial position, while the final clause contain a different overt primary object.

(1.55)

火鉢 に かざし て 、 文字 を あぶり出し た ．

    In (1.56) as well, the zero pronoun subject of the final clause (the main clause) is the antecedent of an empty position in the non-final clause (the precedent of the coordinated clauses), but the zero-pronoun primary object is not. The non-final clause has its own overt primary object. In this case, the zero-pronoun primary object of the final clause must be placed in non-initial position, in violation of the default, after the non-final clause.

(1.56)

「 私たち の 話 を 聞い て 、 監視 し て いる の よ 」

    Following the default practice of placing zero pronouns in clause-initial position could establish an undesired control relation (see section 5.3 for details). Such a situation is caused by the antecedent accessibility hierarchy, a part of the calculation of antecedents for co-reference relations which stipulates that controlled positions are more accessible by object-antecedents than by subject-antecedents. Given the antecedent accessibility hierarchy, subject-antecedent control into a subordinate clause requires that no object appear to the left of that clause. See an example of this scenario below. A clause IP₂ headed by a transitive verb but not containing an overt NP-OB1 argument, shares a subject with a subordinate clause IP₁ , but does not share an object. The more accessible NP-OB1 argument cannot appear to the left of IP₁, lest it bind into IP₁ (i.e., the former be interpreted as the subject of the latter), so it is placed after IP₁.

(1.57)

急い で 食べ ない で 。

    For other configurations, there are fewer conditions on drawing co-reference relations. The clause marked with a conjunctive particle たら in (1.58) contains an overt object PP that co-refers to the null subject of the main clause. Precedence is a sufficient condition for inferring the possibility of co-reference here, and this motivates the non-initial placement of *pro*. Placing the zero pronoun (NP-SBJ *pro*) of the matrix to the right of the IP-ADV conditional clause allows the PP-OB1 argument of the preceding IP-ADV to be taken as antecedent for that zero pronoun, although without some extra information, co-reference is only left open as a possibility, rather than being completely determined. In example (1.58), the PP-OB1 それを of predicate あげ(る)  in the subordinate clause and the null pronoun subject of なくなった in the matrix clause (NP-SBJ *pro*) uniquely share ‘sort information’.

(1.58)

それ を 三 つ 彼 に あげ たら なくなっ た 。

    In (5.68), a subordinate clause and a main clause share both a null subject and a null object. In such a scenario, a null pronoun object cannot be placed at the begining of the main clause, lest it control into the subject position of the subordinate clause. Accordingly, null pronouns sharing ‘sort information’ are introduced into both clauses, with the object of the main clause being placed after the subordinate clause.

(5.68)

ダウンロード し て 印刷 すれ ば 、 学校 や 家庭 で 手軽 に 取り組める 。

1.6   Annotation of core grammatical roles

This section explains how core grammatical roles are annotated in different environments. Section 1.7 will describe how peripheral grammatical roles are annotated.

    First, we address overt instances of NP with core grammatical roles in the simple clause. Core grammatical roles are the roles that nominal arguments take with respect to the predicates that select them. An argument is, in principle, a constituent without which the predicate cannot be fully interpreted. However, there are many instances where it is hard to judge whether a constitutent that co-occurs with a given predicate is obligatory or optional. For example, it is difficult to judge the status of path denoting elements co-ocurring with predicates of motion. As another example, the status of quotational phrases with regard to verbs of cognition and communication may be ambiguous. At present our policy with regard to the first scenario is to mark paths as primary objects when they appear overtly, but to not require null element paths to be annotated for verbs of motion otherwise. Our policy with regard to the second scenario is to treat quotational phrases as having the same status (CP-THT-OB1) with respect to a verb like 言う as to a verb like 話す, but to indicate their adverbial status (CP-THT-ADV) when appearing with verbs that don't involve propositional content at all (see section 6.18.5).

    Core grammatical roles are identified by node label extension -SBJ (subject), -SBJ2 (secondary subject), -LGS (logical subject), -OB1 (primary object), -OB2 (secondary object), -CMPL (complement), and -CZZ (causee). They are typically marked by particles (P-ROLE) が, に, を, の (see section 2.3 for details), etc. However, they may also be marked by toritate particles (P-OPTR) like は and も. They can also appear unaccompanied by particles. This means that the grammatical role of a constituent cannot be reliably predicted solely on the basis of the form of the attaching particle.

1.6.1   Explicitly marked arguments

When a constituent has a core grammatical role, an extension indicating that role is added to the node label. In (1.60), the particle phrase arguments (PP 漁夫が), (PP その女房に) (PP 金を) are marked -SBJ (subject), -OB2 (secondary object), and -OB1 (primary object), respectively.

(1.60)

漁夫 が その 女房 に 金 を 渡し て いる ところ だっ た 。

We assume that the realisation of a constituent with the grammatical role of subject is obligatory for virtually every predicate that heads an IP, so that 1-place predicates have only the subject as argument. Subjects are generally able to serve as antecedents to reflexive 自分 and are able to license honorification morphology on the predicate. Semantically subjects are typically Agents of transitive and unergative predicates, Patients of unaccusative predicates, and Experiencer or Possessor or Attribute-holder of stative predicates.

    As a general principle (noted in section 1.3.6), every IP has a subject, and every subject/predicate pair projects an IP. Sentences that appear not to have subjects are supplied with a null element subject (NP-SBJ *exp*) (see section 1.5). The principle that all IP layers have subjects (either expressed by phrasal categories or determined by control) is not intended to be a claim about Japanese clause structure in general. Requiring fully saturated argument structures simplifies the model and serves as one way to check accuracy of annotation. Using (NP-SBJ *exp*) involves the judgments of native speaking annotators, rendering ‘subjectless sentences’ easily searchable and accessible for evaluation, given that (NP-SBJ *exp*) distingushes them from controllees. Accordingly, an IP that is not a potential control environment (viz. IP-MAT, IP-SUB, IP-REL) always contains a subject. An IP that is potential control environment (viz. IP-ADV, IP-SMC, IP-EMB) does not contain an NP-SBJ if an immediate upstairs argument controls into that IP. See sections 5.3 and 5.4 for examples of the implementation.

    For arguments other than subjects, we refrain from associating their grammatical roles with traditional semantic descriptions. Note that a non-subject argument in a 2-place predicate is formally an -OB1, i.e., the primary object in the sentence. Thus for example a non-subject に-marked argument that might not be regarded as ‘direct object’ in traditional grammar will nevertheless be assigned the label extension -OB1.

(1.61)

散歩 を し て いる うち に 、 来る とも なく 駅 まで 来 て しまっ た 。

    If an argument appearing with a 3-place predicate is (or can be) marked by the particle を, we regularly mark that argument as -OB1. The に-marked beneficiary NP of donative verbs is assigned the grammatical role -OB2.

(1.62)

叔父 は 花子 に 小遣い を 与え た 。

    In the ‘double subject construction’, the second subject (viz. the one closest to the predicate) is frequently marked by が. The constitutent gets the extension -SBJ2 on its node label.

(1.63)

太郎 は 背 が 高い 。

Note: It is not the case that every element marked as -SBJ is actually a subject of the local predicate. In a context where -SBJ precedes and is sister to -SBJ2, in fact -SBJ2 is interpreted as the subject of the predicate that heads the containing clause. An exhaustive search of subjects expressed with nodes would thus require excluding -SBJ when it is sister to -SBJ2, and including -SBJ2 in its place.

    In a passive sentence, に and some other particles mark the semantic Agent of the action denoted by the core predicate, i.e., the ‘logical subject’. In this case, the grammatical role is indicated by the label extension -LGS.

(2.27)

ジョン は 先生 に しから れ た 。

1.6.2   Implicitly marked arguments

NP instances which perform a core grammatical role may not be necessarily marked by a particle associating that role with a predicate. To begin with, they may occur with toritate particles like は and も. In these cases as well, an extension is added to indicate the core grammatical role. In (1.65), the PP headed by は receives the extension -SBJ.

(1.65)

花子 は まだ 学生 に 見える 。

When a particle is omitted, there is no reference to the omitted particle and the NP is linked directly under a label specifying the grammatical role, e.g., NP-SBJ, NP-OB1, etc.

(1.66)

君 、 あの 本 読ん だ 。

1.6.3   Arguments other than NP or PP

There exist predicates selecting arguments which don't have the phrasal category NP or PP, most importantly CP-THT and IP-SMC. They are distinguished from adjuncts (see section 1.7 on adjuncts) by extension tags (-OB1, -SBJ etc.). For detais, see section 6.18 on CP-THT and sections 5.1.4, 4.6, 6.10 and 6.12 on IP-SMC.

    Furthermore, in constructions with verbs する and なる preceded by onomatopoeic or demonstrative adverbs, the entire adverbial phrases serving as complements are marked up as ADVP-CMPL.

(1.67)

ナシャ が イライラ し て 言っ た 。

(1.68)

こう なっ た 以上 、 もう どう し よう も ない 。

1.6.4   Omitted arguments

When an argument with a core grammatical role does not appear explicitly in the sentence, it may be a controllee controlled by an NP in a higher clause, or a gap corresponding to an ATB extracted element. In these cases the ‘gap’ is left empty, without any node. In the case of a relative clause, the ‘gap’ is marked up with a trace as a terminal node: *T*. All other unexpressed arguments are treated as zero pronouns, with terminal nodes in the tree chosen from a closed set: e.g., *pro*, *exp*, *arb*, *hearer*, *speaker*, *hearer+speaker*. These are put directly below a phrasal node with a label showing the grammatical role like NP-SBJ, NP-OB1, NP-OB2, NP-SBJ2, NP-LGS. For details, see section 1.5.

(1.69)

けれども 、 いつ 来る か 、 わから ない 。

    When a subject NP is omitted in a subordinate clause, its reference may be determined by a controller in the matrix clause. In such cases, no corresponding node is added to our annotation. In (1.70), the reference of the subject of the verb embedded by で is provided by the controller, i.e., the subject of the matrix clause わたし.

(1.70)

急い で わたし は 出かけ まし た 。

1.7   Peripheral grammatical roles

Peripheral grammatical roles are performed by adjuncts (adverbial clauses, adverbial phrases, and various kinds of PP). Adjuncts are not essential to the interpretation of the predicate: They are optional constituents.

1.7.1   Explicitly marked adjuncts

Nominal adjuncts marked by particles に, へ, で, から, まで, と, etc. form PP adjuncts. For PP adjuncts, the node labels are desired to have the extensions depending on their semantics such as -LOC (spatial location), -TMP (temporal location), -MSR (temporal extension or frequency), or -ADV (other adverbial function), as the following examples show.

(1.71)

あそこ の ベンチ で 食べ ましょ う 。

(1.72)

つぎ の 汽車 は 七 時 に 出る 。

(1.73)

この 手 の 話 は 何 回 も 聞き まし た

(1.74)

太郎 は 500 メートル も 泳げ た 。

    However, unlike PP arguments with core grammatical roles, PP adjuncts do not always have node label extension in the current annotation. There are many PP instances without extensiton, like PP instances projected by the particles で and に in (1.75).

(1.75)

ビル は 列車 で パリ に 行っ た 。

    There are also particles like に that can mark either an argument or an adjunct. With regard to the obligatory usage of に, see section 1.6.1.

1.7.2   Implicitly marked adjuncts

When NP or PP instances formed without role particles function as adjuncts, depending on their semantics, their node labels can receive the extensions -LOC (spatial location), -TMP (temporal location), -MSR (temporal extension or frequency), or -ADV (other adverbial function).

    In the examples below, the expressions あちこち and この町は denote a spatial location.

(1.76)

両親 は 子供 を ___あちこち 連れ回し た ．

(1.77)

この 町 は いい レストラン が 少ない です 。

    In the examples below, the expression 先週の土曜日 and きのうは indicate a point in time at which an event occurred.

(1.78)

先週 の 土曜日 、 街 で 昔 の 友人 に 会っ た 。

(1.79)

きのう は かぜ を ひき まし た 。

    In the examples below, the expressions ５年間 and 食事中は denote the duration (extension) of an event or state. In the following example, the expression 一度 denotes the frequency (distribution) of an event.

(1.80)

また ５ 年間 使い 続ける ぞい ！

(1.81)

食事中 は テレビ を 見 ませ ん 。

(1.82)

私 は 、 途中 で 一 度 、 悪い 夢 を 見 た 。

For negations and other stative predicates, disambiguation between -TMP and -MSR may seem more difficult at times, but it is useful to remember that the end of the duration denoted by -MSR implies a change of state or end of event, whereas this usually is not the case with -TMP.

(1.83)

今日 は 、 誰 に も 会わ ない 。

If a combination of expressions ‘measure out’ an event, the expressions are treated as instances of NP-MSR:

    In the examples below, the expression 予定通り and 本当は have an adverbial function which does not correspond to any of the more specific categories.

(1.84)

会議 は 、 予定通り 3 時 に 始まっ た 。

(1.85)

本当 は 叔母 の アイリーン です が

    If a particle is omitted from an NP adjunct, a PP projection is created and the expected particle is inserted as a terminal node that is placed between instances of ‘*’.

(1.86)

それ 、 おれ 行く ん だ けど さー

See also section 2.14.

1.7.3   Adjunct traces

Aside from an ATB environment a ‘gap’ in a relative clause is the only place where an unindexed null element can be associated with an adjunct role. If the adjunct role is temporal or locative, this information can be added as an extension of the label. For example, (1.87) has trace: (NP-LOC *T*).

(1.87)

ここ が 高津さん が 講演 し た ところ だ 。

    If the adjunct role associated with the modified head N is neither temporal nor locative and usually indicated by a particle, the trace can be located in an NP under a PP with the missing particle specified in a terminal beginning and ending with ‘*’. For example, (1.88) has trace: (PP (NP *T*) (P *に*)).

(1.88)

早朝 、 彼 は 住み慣れ た 町 を 出で立っ た ．

1.8   Antecedent relations

Sort information is used for resolving reference for pronouns. Pronouns such as これ、あいつ、そこ, etc., and the null pronoun *pro* trigger the need to introduce sort information that establishes a link with the accessible antecedent.

    If an NP X (either nominal or pronominal) both c-commands and shares sort information with a pronoun Y, and there is no closer candidate to Y which bears the same sort information, an antecedent relation will be calculated between X and Y.

    If a nominal NP X precedes pronoun Y in the discourse, and shares sort information with Y, and there is no closer candidate to Y which bears the same sort information, an antecedent relation will be calculated between X and Y regardless of their respective structural positions.

    One exception is when an indefinite element is inside of an ‘operation of closure’ such as the scope of negation or modality. In such environments, only definite elements can serve as antecedents to outside pronouns.

    Note that an identity relation obtaining between two nominal expressions will not be calculated through sort information. Such information can be safely added, if desired, and such information is potentially useful for linguistic studies and a full representation of the semantics of texts, but it is outside the purview of Scope Control Theory. Cataphoric relations are also exceptions in this sense, but are interesting and worth marking up.

1.8.1   Sort information

This section details the situations where sort information is needed.

1.8.1.1   Non-c-commanding antecedents

(1.89)

しかし 、 その 内容 は 『 タイガーマスク 』 に 言及 し 、 「 自分 も 仮面 を 被れ ば ヒーロー に なれる 」 という 児童 の 願望 を 指摘 し て いる こと 、 主人公・九条剛 が 普通 の 体育教師 で 鍛錬 によって ヒーロー の 力 を 得 て いる など 、 当時 流行 し て い た スポーツ根性もの の 影響 が 強い 内容 で あっ た 。

1.8.1.2   Islands

While IP-SMC under CP-THT allows control from an immediately higher IP, IP-SUB under CP-THT is an island for control. In (1.90) below a c-commanding pronoun becomes an antecedent for another pronoun in a quotation by virtue of sort information.

(1.90)

彼 は たびたび 僕 に 「 私 は もう 長く は ない 」 と 言い 、 僕 は それ を 戯言 と 思っ て い た 。